Skip to content

Bioinformatics Lab UI

Goals served: all goals (interactive exploration interface)

Tier: Mature

Last updated: 2026-04

Problem

Much of splice-prediction research is exploratory: comparing base-model predictions for a gene, inspecting a chromosome-wide distribution, cross-referencing metrics across training runs. A CLI or notebook is inefficient for these tasks. Domain scientists benefit from an interactive web interface with gene browsing, on-demand prediction, and model-comparison dashboards — all backed by the same src/ library as the rest of the system.

User-facing functionality

  • Browse and search ~19K genes with filters (chromosome, biotype, length)
  • Request on-demand splice prediction for any gene with per-model switching (SpliceAI vs OpenSpliceAI)
  • Visualize 3-track Plotly genome view (annotations, predictions, delta)
  • Compare evaluation runs across models and thresholds in a metrics dashboard
  • Peak-preserving downsampling for genes > 10K positions
  • LRU-cached predictions keyed by (gene, model)

Driving examples

Start the server:

mamba run -n agentic-spliceai python -m server.bio.app
# Browse: http://localhost:8005/

Pages:

  • / — Gene Browser
  • /genome/{gene} — Genome View with on-demand prediction
  • /metrics — Metrics Dashboard

API endpoints:

Gene browser + genome view:

  • /api/genes, /api/genes/stats, /api/genes/chromosomes
  • /api/genome/{gene}/predict?model=X&threshold=T
  • /api/metrics/runs, /api/metrics/{run_id}, /api/metrics/compare

Ingestion-layer readiness (Phase D3 — read-only wrappers over the data_preparation and multimodal_features applications' get_status() APIs; see server/bio/ingest_api.py):

  • /api/ingest/health — liveness check
  • /api/ingest/data-prep/builds — list configured base-model / build entries
  • /api/ingest/data-prep/status?build=GRCh38&annotation_source=mane — canonical-path check
  • /api/ingest/data-prep/status?output_dir=/path — explicit output-dir check
  • /api/ingest/features/profiles — feature-profile catalog (default, full_stack, …)
  • /api/ingest/features/tracks?build=GRCh38&modality=conservation — external-track catalog (UCSC URLs, ENCODE accessions)
  • /api/ingest/features/status?build=GRCh38&chromosomes=1,2,21,22 — per-chromosome feature-parquet readiness

All ingestion endpoints are read-only. Writes (running prepare) remain CLI-only; a future async-job endpoint would be a separate router.

src/ surface

  • server.bio.app — FastAPI app entry
  • server.bio.bio_service — prediction cache, service logic
  • server.bio.config — MAX_CACHED_PREDICTIONS, color palette, etc.
  • agentic_spliceai.splice_engine.base_layer.* — prediction backend
  • agentic_spliceai.splice_engine.data.preparation — annotation extraction
  • agentic_spliceai.splice_engine.eval.* — metrics aggregation

Design patterns (see CLAUDE.md for details):

  • LRU prediction cache (OrderedDict in bio_service.py), keyed by (gene, model), threshold only affects classification (raw predictions cached)
  • Peak-preserving downsampling for sparse peaky data (never naive [::factor])
  • reload_dirs=["server/bio"] in app.py prevents data writes from triggering uvicorn restart

Evaluation

  • Genes browsed: ~19K (filtered to canonical chromosomes)
  • Demo notebook: SERPINA1 (COPD), UNC13A/STMN2 (ALS) clinical use cases
  • Throughput: prediction cache hit rate dominates interactive workflows
  • No formal benchmark — quality measured by interactive usability

Maturity tier and signals

Current tier: Mature

Signals supporting the tier:

  • Running FastAPI service with stable endpoint contracts
  • Demo notebook demonstrating end-to-end clinical use cases
  • LRU cache + downsampling patterns documented and tested in practice
  • Used as the primary exploratory interface across sessions
  • Phase 2.5 marked complete in ROADMAP.md

Graduation signals

To advance to Product, the application needs:

  • Deployment guide (containerization, reverse proxy, auth)
  • External hosting decision (intranet only today)
  • Versioned API contract with deprecation policy
  • Inference-path tests covering all endpoints
  • Rate limiting / resource quotas for on-demand prediction

Known limitations

  • Single-process service; on-demand prediction blocks the event loop during model inference
  • No authentication or access control — intended for local/dev use
  • Prediction cache is in-memory only — cold start loses cache on restart
  • No multi-user session support; threshold changes affect all users
  • Plotly rendering for very large genes relies on downsampling; exact per-position view requires dedicated endpoint